Button, watch, sensor or camera: what each does well and where it breaks

Four technologies are sold for the same job: making sure someone finds out if your mother or father falls while alone. The pendant button, the watch with automatic detection, motion sensors, and cameras.

None of them wins. Each solves part of the problem well and fails at another part, and which one suits you depends on fairly untechnical things: whether the person will wear something, how the home is laid out, and who is at the other end to act on the alert.

None of them prevents the fall, either: they all try to shorten time on the floor, the minutes or hours between someone falling and someone finding out. And they all fail in the same two ways — not alerting when something happened, and alerting when nothing happened. The first you pay in full, on the night it mattered; the second costs slowly and therefore fools people: three false alarms and the family turns the volume down. A switched-off system detects nothing.

The pendant button

It's the only one that interprets nothing: if the button was pressed, a person decided to ask for help. It works in the dark, it works in the bathroom, and someone usually answers at the other end.

It fails because it needs two conditions at once: being worn, and being pressable. The first breaks on its own: many people take it off to sleep and to shower. The second breaks in the worst case: a fall that leaves someone dazed or with an arm pinned isn't solved by pressing anything. And there's a failure that isn't technical: plenty of people don't press it even when they can, because they don't want to bother anyone at three in the morning.

The watch with automatic detection

It solves half of that: the pressing. The accelerometer detects the impact and the alert goes out by itself. It's worn on the body, so it also covers outside the house, where no sensor or camera reaches.

But it still has to be worn, and it has to be charged — and the charging window usually lands at night, which is when it's needed most. The mechanism has its own limit: an accelerometer detects impacts, not falls. A slow fall, like sliding off a chair until you end up sitting on the floor, may never produce the spike the algorithm waits for; a hard knock of the wrist against a table, on the other hand, looks a lot like one.

Motion sensors

Nothing is worn. Nobody has to remember or charge anything, and there's no image of anyone anywhere: it's the option that asks the least of the person being cared for.

But they don't see people, they see movement. They can't tell "asleep" from "on the floor": they only know nothing has moved for a while. They work by absence, and the alert arrives once the routine is already broken — hours, not minutes. They're slow by design for a fall. They're also easy to confuse: a visitor or a cat is enough to make everything look normal.

Cameras

It's the only category that can tell "not moving" from "on the floor", because it looks at the scene instead of inferring it. On this one we have our own numbers:

• No light, nothing at all. We left a phone watching a dark room for a whole night: 540 minutes, peak brightness 0 on a 0-255 scale, zero people detected. It isn't the model — with no light there's no image to analyse.
• Static posture says less than you'd think. In our measurements, a standing person is detected with a median confidence of 0.702 and appears in 97.8% of frames; on the floor that drops to 0.339 and 11.8%.
• Sitting on the floor and lying on the floor look too alike. It cost us two false alarms with someone sitting on a beanbag: the silhouette is nearly identical, and what separates them isn't the posture but what happened immediately before.
• It only sees what's in frame. The hallway and the bathroom don't exist to a camera.

And there's the objection that isn't technical: it's a camera inside your mother's home. The question is where the video is processed and what leaves the house.

How to choose when there's no winner

• What does it ask of the person being cared for? Anything to remember, charge or wear will fail some day.
• What does it do at three in the morning with the lights off? A useful answer is a measurement, not an adjective from a brochure.
• What happens when it gets it wrong? How a false alarm is cancelled, and what the family hears in the meantime.
• Who receives the alert, how long does it take, and what do they do next? An alert with nobody at the other end isn't detection.

And something rarely said out loud: these aren't mutually exclusive. What's worn covers the street; what's fixed in the house covers the hours when nothing is being worn.

What we do, and what we don't

Velia sits in the fourth category. It turns an Android phone you already own into a care aid that analyses images on the phone and provides best-effort alerts of possible falls. It does not store video; live video travels encrypted only when an authorised person opens it. It asks nothing of the cared person: nothing to wear, nothing to press, nothing to charge at night.

It inherits its category's flaws, so we say them here and inside the app: without light it doesn't see, and when that happens it says so on screen instead of showing that all is well. It alerts your family, not a monitoring centre staffed around the clock, and it isn't a replacement for the emergency button of someone who already has one and uses it.

Choosing well means knowing what fails. That's why we publish ours.